AI enters construction through three doors. Site back office, quote processing, then site data. The first two produce a measurable result within weeks. The third takes a year of work and a tidy document base.
This article gives the figures from the sector study published in January 2026. It shows what Eiffage and VINCI Construction actually deploy. It then details two engagements Tandem ran, one in an estimating department, the other at a joinery manufacturer. Every figure carries its source and its date.
Where AI in construction really stands in 2026
The sector is barely starting. France's Observatoire des métiers du BTP commissioned consultancy Plein Sens to run a study on the perception and integration of AI tools in construction companies, published in January 2026. It rests on 621 company leaders who completed the questionnaire, out of 45,000 professionals contacted. The result is blunt.
- 3% of companies have deployed AI solutions.
- 5% are currently deploying.
- 36% are considering it in the coming years.
- 40% do not use it and have nothing planned.
Personal usage is no further ahead. 43.5% of respondents say they have never used a consumer tool such as ChatGPT, Copilot or Gemini. The gap between company sizes is brutal. Small firms cap at 7.3% current usage, while close to 35% of large companies have already deployed or are deploying.
The study also cites a Bpifrance Le Lab survey of more than 1,200 leaders of French SMEs and mid-caps. 58% of them see AI as a strategic priority in the short or medium term. Only a third have launched a project, and in 73% of cases the leader personally drives those first uses. A topic carried by one person stays fragile.
What Eiffage and VINCI Construction already deploy
Large groups act as the laboratory, and their figures are public. Eiffage launched its generative AI training in February 2025. In the case study published by Google Cloud in April 2025, the group reports 2,500 employees enrolled and 2,000 trained. 1,600 actively use the tools, out of 52,000 people in France. Fifteen use cases were selected. The most telling one covers tender pack analysis, cut from two days to twenty minutes.
VINCI Construction published an update on its in-house tools on 13 April 2026. Its Select tool analyses tenders and brings that analysis from several hours down to about thirty seconds. Caméléon, built by the Sixense teams since 2021, automatically flags structural anomalies on drone imagery.
Both examples share one trait that matters to an SME. Neither touches the act of building. They work on documents, images and files. That is exactly the material every construction company already produces, with no sensor and no infrastructure investment.
Quotes, the first seam in an estimating department
Tandem ran a multi-week audit inside the pricing and estimating department of the Coulon group. The company has been doing facade renovation, insulation and waterproofing since 1920, with around 120 people in Vanves. The findings match most estimating departments. Hundreds of quote lines retyped by hand into every architect schedule, project after project. Dozens of emails per day per site manager, sorted with no shared rule across business units. And eighteen years of accumulated documents, scattered and out of reach for any AI.
The audit output is not a shopping list. The selected cases are sorted into three modes, based on how much human validation they need. Autopilot for repetitive volume, Copilot for high-stakes tasks, Hybrid when the process mixes both. Estimated savings exceed a thousand hours per year. One case sums it up, a 200-page technical specification becomes searchable in seconds.
The document foundation stays the prerequisite. Without a shared folder structure and a migration of project files, most selected cases collapse. This is the point companies underestimate the most.
How a PDF quote becomes an ERP line
The second ground is industrial. FPV Industries manufactures PVC, aluminium and timber joinery, with multi-leaf products and conditional options. Every quote arrived as a PDF and left as manual entry into the in-house ERP. Someone had to read it, interpret it, match the product to the catalogue, link each option to the right integration tag, check colour consistency, then produce the expected file. A grid of nearly a thousand business rules, applied quote by quote.
The workflow Tandem built triggers on intake through a webhook. An OCR engine reads the PDF, a chain of models detects products and dimensions, then the workflow matches the catalogue and maps the options. It outputs a custom JSON sent to the ERP web service. Thirty seconds of processing, around 15 cents per quote, zero re-entry on the covered scope.
OCR first, the model second
The technical detail that decides the outcome fits in one sentence. A large language model is not the best PDF reader. I explained this in a post about extracting data from PDFs, in April 2025. At a previous employer, quarterly reports from asset managers changed format with every batch, with numbers sitting inside images. Raw extraction returned the wrong data. Home-made scripts broke on the first layout change. Manual entry held up, but slowly.
The setup that works chains two steps. An OCR engine first produces a machine-readable version, text, structure and layout included. A model then analyses that output with a prompt that forces verification steps, and returns JSON with the expected variables. No template left to maintain, even when fields change order. The same pattern covers a supplier quote, an invoice or a purchase order.
Processing cost drops very low. In my newsletter issue on AI use cases in business, I place document extraction under 5 cents per document on standard cases, orchestration included. The gap with human entry is not up for debate. The real question becomes the escalation rule, what runs automatically and what goes back to a human.
Site back office, the fastest win
This is the least spectacular workstream and the most profitable one. Site managers spend their days on site and their evenings at the desk. The Observatoire study documents a practice that is already widespread, often outside any official framework. Site managers record themselves during progress meetings and inspection visits, then have the summary and the emails drafted.
It frees up their time. They can spend more of it on control and on managing teams, and drop everything they dislike, the paperwork. Report quality goes up, and so does the quality of how information circulates.
That quote comes from the head of an SME building individual homes, cited in the January 2026 study. The effect spreads further than expected. Administrative teams used to spend a large share of their time correcting documents written by site foremen before filing them.
One deadline makes this workstream urgent. Electronic invoicing becomes mandatory in France on 1 September 2026. From that date, every company must be able to receive an electronic invoice, and large companies and mid-caps must issue them. SMEs, small firms and micro-businesses follow on 1 September 2027. A normalised, structured invoice flow makes accounting automation markedly easier, something accounting firms already see in their own AI use cases.
Why AI projects stall in construction
Three obstacles come back, and none of them is about the model. The first is interoperability. Estimating, BIM, site tracking and invoicing software barely talk to each other. The study quotes a technical director at a civil engineering mid-cap who says he spends 70% of his time building bridges between his tools.
The second is the data itself. A model does nothing useful with eighteen years of files scattered across servers, mailboxes and photo galleries. Tidying before automating is not a preparation step, it is the project.
The third is economic. With 94% craft businesses in the building trades and thin margins, a project whose return is not visible within a quarter simply does not start. Hence the simple rule below.
| Use case | What it replaces | First result | Prerequisite |
|---|---|---|---|
| Dictated site visit report | Evening data entry at the desk | 2 weeks | A smartphone and a written usage rule |
| Site email triage | Manual sorting by the site manager | 1 month | A folder structure shared across teams |
| Search inside a technical spec | Skim-reading the tender documents | 1 month | Technical documents gathered in one place |
| Reading supplier quotes | Line-by-line retyping | 2 to 3 months | A catalogue and written business rules |
| Progress tracking from images | The weekly visual survey | 6 months or more | Regular, consistently framed photography |
Where to start with AI in construction
Start with the back office, always. A dictated site report and a sorted mailbox go live within weeks, without touching your core systems. The gain shows up in the first few days. Tackle quotes next, because they require a clean catalogue and written rules. Keep site data for the following year. That is the order Tandem applies across its construction and industry engagements, and it barely varies from one company to the next.
Refuse any project without a measurement point. A use case with no known volume and no baseline duration will produce a successful demo and no decision behind it. That is the role of an AI audit, quantify before building, then ship a first brick within weeks using an n8n workflow. The full list of the cases that come up most often sits in our panorama of AI use cases in business.
